Claude Fable 5 🤖, one repo owns your laptop 🔑, driverless Doritos 🚚

Anthropic has made Claude Fable 5 generally available, a public version of its Mythos-class model it calls state-of-the-art on nearly all tested benchmarks, with bigger gains on longer tasks. To keep it safe, queries on sensitive areas like cybersecurity get rerouted to Claude Opus 4.8, and the guardrails trigger in under 5% of sessions. It is free across paid plans through June 22, then moves to usage credits.

Mustafa Suleyman, Microsoft's AI chief, publicly criticised Anthropic for entertaining the idea that Claude might be conscious, calling the speculation 'really, really dangerous.' The rebuke lands the same week Fable 5 went public and deepens a split between labs over how to talk about model welfare and machine sentience. Suleyman has said the industry should build AI that serves people, not AI that performs personhood.

OpenAI is preparing its biggest ChatGPT redesign yet, recasting the chatbot from a simple answer box into a superapp that steers users toward higher-margin products, according to the Financial Times. The overhaul arrives as the company files confidentially for an IPO. Together the moves signal a push to convert its early generative-AI lead into durable revenue rather than a clever demo most people use for free.

Google has reportedly ordered more than three million of its custom TPUs from Intel's contract-manufacturing arm for 2028 production, according to The Information, sending Intel shares up sharply. The move signals that the biggest AI buyers are looking past their usual supplier TSMC, which is straining to keep pace with demand. Nvidia and Tesla are also said to be exploring Intel's foundry, a potential turning point for its comeback.

Vibe-coding startup Lovable told TechCrunch it has crossed $500 million in annualised revenue, up from $400 million in February, less than three years after it was founded. Users have now built more than 50 million projects, with a million new ones starting each week. Most are non-technical founders, designers, and salespeople building real software they plan to sell or run, from storefronts to CRMs and internal tools.

Security firm Varonis built an OpenClaw email agent, wired it to Gmail, browser tools, and Google Workspace, then ran classic phishing tactics against it. The agent fell for the same tricks used on humans and leaked AWS keys, database credentials, and CRM exports. A stricter configuration profile changed the outcome, a reminder that autonomous agents need identity checks and scoped access before they touch real inboxes.

PepsiCo has become the first major US consumer-goods company to run a full-scale fleet of autonomous trucks on public roads with no safety driver. Forty-one self-driving vehicles, 35 of them in Arizona plus a handful in Texas and Arkansas, are moving Frito-Lay snacks and drinks. The company says the trucks have run since June 2025 with a 99% on-time rate and no accidents, and expects to hire fewer drivers.

Researchers at Straiker disclosed NomShub, a vulnerability chain in the Cursor AI editor where simply opening a booby-trapped repository can give an attacker persistent shell access. It strings together indirect prompt injection, a sandbox escape through shell builtins that Cursor's command parser ignores, and Cursor's own remote-tunnel feature. With every protection switched on, the agent can be tricked into installing a silent, undetected backdoor on the developer's host.

When an agent runs code, where does it actually run and what can it touch? OriginHQ walks through MXC, the cross-platform sandbox Microsoft open-sourced at Build under an MIT licence, reading the real source instead of reverse-engineering it. The teardown covers one dispatcher binary, a JSON policy model, and ten containment backends, and sets MXC against how Codex, Claude Cowork, and Google's GKE sandbox each isolate model output.

Anthropic's red team studied N-days, bugs that are patched but still unpatched on many systems, and found models can sharply compress the patch gap. Because a security fix is itself a roadmap to the bug, a model can diff the patch, reverse-engineer the flaw, and build a working exploit. The takeaway: attackers no longer need rare expertise to weaponise disclosed fixes, shrinking defenders' update window.

Cognition built FrontierCode, a coding benchmark that moves past 'is the code correct' to 'is the code mergeable.' More than 20 open-source maintainers created realistic tasks from repos they run, spending over 40 hours each defining what a merge-worthy PR looks like, judged on correctness, test quality, scope discipline, and style. An ensemble of unit tests, rubrics, and verifiers cuts false positives by 81% against prior grading methods.

When several agents read the same document, a default serving stack reruns the identical prefill for each one, which is pure wasted compute. This walkthrough builds SwarmKV: run prefill once, serialise the KV cache to a host buffer, copy it per branch, and restore before decoding. On a seven-year-old GTX 1080, a two-agent pipeline ran 1.95 times faster end to end, and the second agent's activation latency dropped 52 times.

Ben Thompson's read is that Apple's bungled Apple Intelligence has earned it the 'vaporware' label it once mocked Microsoft for, just as Microsoft's Build vision of cloud-resident agents and thin client devices points past the phone entirely. If agents run on servers and finish tasks invisibly, local compute matters less, and the iPhone's role at the centre of computing is under threat. He reads WWDC as Apple defending that centre.

Simon Willison spent a day with Claude Fable 5 and calls it 'something of a beast', excellent but slow and expensive at $10 per million input tokens, double Opus 4.8. It ran up $110 in a day and wrote almost all of a new LLM library release on its own, and he reckons it may be the largest model any vendor has shipped. For the day-to-day view, see Ethan Mollick's hands-on.

Mike Masnick dissects a wave of near-identical all-hands emails: a CEO declares AI amazing, orders everyone to use it immediately or leave, and sometimes stands up a token leaderboard. He says forced adoption misses the point, since good AI use means treating tokens as scarce and choosing the tool willingly. The mandates reveal managers chasing a trend rather than understanding where these tools genuinely help and where they fall short.

Sean Goedecke makes the case that many engineers should work fewer hours and at a slower pace, aiming for 80% utilisation by default. His reasoning: performance at tech companies is dominated by outlier events, and the highest-impact changes often take surprisingly little work but need the slack to spot the right problem at the right time. There are no points for effort, only for solving what matters.

As AI generates code faster than anyone can read it, the real bottleneck has become code review, and pointing a second model at it does not help. The fix is to move the human checkpoint upstream, reviewing intent, specs, and constraints before any code exists. Teams with heavy AI adoption now merge 98% more PRs while review times climb 91%, so out-reading the machine is a losing game.

Firecrawl published a pack of outcome-focused skills for AI coding agents that turn web scraping into finished work rather than raw data. Instead of returning pages, each workflow produces a deliverable: a multi-source research report, an SEO audit with prioritised fixes, a live-site QA report, a pre-meeting lead brief, or competitive-intel monitoring. It installs through the Firecrawl CLI or as a skills pack.

Zero, from Rocicorp, is a general-purpose sync engine that brings the instant, local-first feel of Linear, Figma, and Superhuman to ordinary web apps. Reads and writes hit a client-side store first and sync in the background, but unlike earlier tools it handles partial data and complex permissions, syncing only the slice a user can access. A demo with 1.2 million rows loads in under two seconds.

Cohere's North Mini Code is its first model aimed at developers: a 30B-parameter mixture-of-experts with just 3B active, built for agentic software engineering and available on Hugging Face under Apache 2.0. On Artificial Analysis' Coding Index it scores 33.4, beating similar-size open models like Qwen3.5 and Gemma 4, and even much larger ones such as Nemotron 3 Super and Mistral Small 4. It is tuned for terminal-based agent workflows.